Instance-Consistent Fair Face Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Fairness in face recognition (FR) is a significant challenge for current algorithms in diverse societies.
• Existing fair FR methods often overlook the misalignment between training and testing metrics.
• Ensuring equitable performance across different demographic groups is crucial for ethical AI deployment.

Purpose of the Study:

• To propose a novel instance-consistent fair face recognition (IC-FFR) method.
• To achieve complete instance fairness concerning false positive rates (FPR) and true positive rates (TPR).
• To address the metric misalignment issue in training and testing phases of FR algorithms.

Main Methods:

• Theoretical analysis correlating testing metrics (FPR, TPR) with label classification loss (softmax loss).
• Development of an instance-consistent fairness solution using customized instance margins.
• Introduction of the National Faces in the World (NFW) dataset for fine-grained fairness evaluation.

Main Results:

• The IC-FFR method demonstrates high-probability consistency of unfairness penalties from FPR and TPR to softmax loss.
• Customized instance margins effectively preserve consistent FPR and TPR for all instances during training.
• Experiments on NFW, RFW, and BFW benchmarks show the method's effectiveness and superiority over state-of-the-art approaches.

Conclusions:

• The proposed IC-FFR method successfully achieves instance fairness in face recognition.
• The theoretical analysis provides a foundation for understanding and mitigating fairness issues.
• The NFW dataset facilitates more rigorous evaluation of fairness in face recognition systems.